Neural Constituency Parsing of Speech Transcripts

Proceedings of NAACL-HLT 2019, pages 2756–2765 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 2756 Neural Constituency Parsing of Speech Transcripts Paria Jamshid Lou Macquarie University Sydney, Australia Yufei Wang Macquarie University Sydney, Australia {paria.jamshid-lou,yufei.wang}@hdr.mq.edu.au Mark Johnson Macquarie University Sydney, Australia mark.johnson@mq.edu.au Abstract This paper studies the performance of a neu- ral self-attentive parser on transcribed speech. Speech presents parsing challenges that do not appear in written text, such as the lack of punctuation and the presence of speech dis- fluencies (including filled pauses, repetitions, corrections, etc.). Disfluencies are especially problematic for conventional syntactic parsers, which typically fail to find any EDITED dis- fluency nodes at all. This motivated the de- velopment of special disfluency detection sys- tems, and special mechanisms added to parsers specifically to handle disfluencies. However, we show here that neural parsers can find EDITED disfluency nodes, and the best neu- ral parsers find them with an accuracy sur- passing that of specialized disfluency detection systems, thus making these specialized mech- anisms unnecessary. This paper also investi- gates a modified loss function that puts more weight on EDITED nodes. It also describes tree-transformations that simplify the disflu- ency detection task by providing alternative encodings of disfluencies and syntactic infor- mation. 1 Introduction While a great deal of effort has been expended on parsing written text, parsing speech (either tran- scribed or ASR output) has received less atten- tion. Parsing speech is important because speech is the easiest and most natural means of com- munication, it is increasingly used as an input modality in human-computer interactions. Speech presents parsing challenges that do not appear in written text, such as the lack of punctuation and sentence boundaries, speech recognition errors and the presence of speech disfluencies (including filled pauses, repetitions, corrections, etc.) (Kahn et al., 2005). Of the major challenges associated with transcribed speech, we focus here on speech disfluencies, which are frequent in spontaneous speech. Disfluencies include filled pauses (“um”, “uh”), parenthetical asides (”you know”, ”I mean”), interjections (“well”, ”like”) and partial words (“wou-”, “oper-”). One type of disfluency which is especially problematic for conventional syntac- tic parsers are speech repairs. Following the anal- ysis of Shriberg (1994), a speech repair consists of three main parts; the reparandum, the interregnum and the repair. As illustrated in the following ex- ample, the reparandum we don’t is the part of the utterance that is replaced or repaired, the interreg- num uh I mean (which consists of a filled pause uh and a discourse marker I mean) is an optional part of the disfluency, and the repair a lot of states don’t replaces the reparandum. The fluent version is obtained by removing the reparandum and the interregnum. �reparandum �� � We don’t �interregnum �� � uh I mean repair � �� a lot of states don’t� have capital punishment. (1) In the Switchboard treebank corpus (Mitchell et al., 1999) the reparanda, filled pauses and dis- course markers are dominated by EDITED, INTJ and PRN nodes, respectively …

Content truncated to 500 words